Understanding how complex human, environmental, demographic and operational factors interact to elevate the probability of safety infractions is essential for developing predictive safety systems in high-risk environments such as highways. This study proposes a data-driven framework for identifying empirical inflection and threshold operational points, whereby risk shifts abruptly from acceptable to hazardous levels. In addition, this study develops interpretable, rule-based triggers for proactive safety interventions in highway maintenance and traffic management.
Using a synthetic dataset (backed by a positivist philosophical stance) reflecting safety-based variables recommended in the literature (e.g. human, operational, environmental stressors and organisational conditions), a supervised machine learning model (i.e. Random Forest) was trained to estimate infraction probabilities. A threshold discovery algorithm was then implemented, combining bin-wise probability estimation with prominence-based inflection detection and rule induction to extract applicable safety triggers. Feature importance measures were used to contextualise the relative influence of predictors on the model's risk output.
Results revealed clear, interpretable thresholds across multiple predictors, including sharp risk transitions for consecutive workdays (3–4 days), fatigue level (=5), sleep duration (<6 h), traffic density (>600 vehicles/hour) and physiological stress (>90 bpm). Non-linear variables such as cognitive load and training quality exhibited oscillatory risk patterns yet still produced meaningful inflection points. Based on the discovered thresholds, safety trigger rules were formulated to aid in safety management decision-making.
This study contributes a novel, transparent methodology for threshold-based risk detection, bridging machine learning interpretability with practical safety management. The proof-of-concept model developed demonstrates how inflection-point analytics can support early warning systems, personalised interventions and data-driven policy design in safety-critical operational settings.
